Loki: Representation over Architecture for Diffusion-Based Portrait Animation
Researchers have developed Loki, a new diffusion-based method for animating portraits that separates identity from expression and pose. Unlike previous methods that struggle with disentangling these factors from RGB data, Loki uses a specialized face model to encode expression and pose, which are then rasterized into a spatial map. This approach significantly reduces the need for cross-identity training data and requires fewer inference parameters compared to existing techniques. Loki also demonstrates leading performance on metrics measuring adherence to driver expression and head pose. AI
IMPACT This new method could enable more efficient and realistic AI-driven portrait animation by simplifying the disentanglement of identity, expression, and pose.
